[Paper Review] LNDb: A Lung Nodule Database on Computed Tomography

João Pedrosa, Guilherme Aresta+9 more|arXiv (Cornell University)|Nov 19, 2019

Lung Cancer Diagnosis and TreatmentMedicine22 references18 citations

TL;DR

This paper introduces LNDb, a new lung nodule database from low-dose CT scans designed to reflect real-world clinical variability and support the development of computer-aided diagnosis (CAD) systems. It evaluates state-of-the-art deep learning methods for nodule detection, segmentation, and characterization, showing they can match radiologist performance in follow-up recommendations—though detection remains a key challenge, especially when adapted to local imaging settings.

ABSTRACT

Lung cancer is the deadliest type of cancer worldwide and late detection is the major factor for the low survival rate of patients. Low dose computed tomography has been suggested as a potential screening tool but manual screening is costly, time-consuming and prone to variability. This has fueled the development of automatic methods for the detection, segmentation and characterisation of pulmonary nodules but its application to clinical routine is challenging. In this study, a new database for the development and testing of pulmonary nodule computer-aided strategies is presented which intends to complement current databases by giving additional focus to radiologist variability and local clinical reality. State-of-the-art nodule detection, segmentation and characterization methods are tested and compared to manual annotations as well as collaborative strategies combining multiple radiologists and radiologists and computer-aided systems. It is shown that state-of-the-art methodologies can determine a patient's follow-up recommendation as accurately as a radiologist, though the nodule detection method used shows decreased performance in this database.

Motivation & Objective

• To address the gap in existing databases by creating a clinically representative lung nodule dataset reflecting real-world radiologist variability and local imaging conditions.
• To evaluate state-of-the-art deep learning methods for pulmonary nodule detection, segmentation, and characterization in a real-world clinical context.
• To investigate collaborative strategies between radiologists and CAD systems to improve detection sensitivity without increasing clinician workload.
• To assess whether CAD systems can match radiologist performance in classifying patients according to Fleischner guidelines for follow-up.

Proposed method

• The LNDb database comprises 294 low-dose CT scans from a single institution, with nodule annotations from two radiologists using a single-blind protocol.
• Eyetracking data was recorded during radiologist reading sessions to capture gaze patterns and attention distribution across nodules.
• State-of-the-art deep learning models were trained and evaluated on nodule detection, segmentation, and malignancy characterization tasks.
• Collaborative strategies were tested by combining radiologist readings with CAD outputs, using both fixed thresholds and dynamic attention-based filtering.
• Performance was evaluated using standard metrics: sensitivity, false positives per scan, and weighted Cohen’s kappa for scan-wise agreement.
• The Fleischner guidelines were used as a clinical benchmark to assess patient follow-up recommendations.

Experimental results

Research questions

• RQ1 Can state-of-the-art CAD systems achieve detection and classification performance comparable to radiologists in a real clinical setting?
• RQ2 How does radiologist variability affect the performance of CAD systems, and can eye-tracking data improve collaboration between radiologists and CAD?
• RQ3 To what extent do collaborative strategies between radiologists and CAD systems improve detection sensitivity without increasing time burden?
• RQ4 Does CAD performance degrade when applied to a local dataset with different acquisition protocols compared to benchmark datasets like LIDC-IDRI?
• RQ5 Can CAD systems accurately classify patients according to Fleischner follow-up guidelines, matching radiologist decisions?

Key findings

• State-of-the-art CAD systems achieved patient follow-up classification accuracy comparable to individual radiologists, demonstrating strong clinical relevance.
• Nodule detection performance was lower in LNDb than in benchmark datasets, indicating that local imaging characteristics significantly affect model generalization.
• Collaborative strategies combining radiologists and CAD improved detection sensitivity, particularly when using attention-based filtering, though gains were marginal due to small sample size.
• Eye-tracking data revealed meaningful gaze patterns that could inform future CAD design, especially for prioritizing regions of interest.
• Despite improvements in sensitivity, collaborative strategies did not significantly outperform individual radiologists or CAD alone in Fleischner follow-up classification.
• The study highlights that model training on local data is crucial for optimal performance, as generalization across institutions remains a challenge.